[Feat]Add new claude 4-6 feat for v1/messages

[Sameerlite]

Relevant issues

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🆕 New Feature
🐛 Bug Fix

Changes

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 9, 2026 11:51am

[greptile-apps]

Greptile Overview

Greptile Summary

This PR extends LiteLLM’s /v1/messages support for newer Claude Opus 4.6 features across multiple providers.

Key changes:

• Azure AI Anthropic: filters anthropic-beta to remove unsupported beta flags after building the beta header set.
• Bedrock Invoke: replaces an inline beta whitelist with a centralized filter_and_transform_beta_headers(provider='bedrock') helper.
• Vertex AI partner models: adds automatic anthropic-beta injection for context management (context-management-2025-06-27) and compaction (compact-2026-01-12) based on optional_params['context_management'].edits, with new unit tests.
• Pricing JSON: removes duplicated keys for claude-opus-4-6-v1 variants, which previously would have caused silent overwrites in JSON parsers.
• Docs: adds a /v1/messages usage section and feature matrix for Opus 4.6.

The main correctness concern is in the Vertex AI context management detection: it can add the context-management beta header even when edits are empty/invalid (e.g., {} or missing type), and the docs currently claim compaction support for Bedrock Invoke which conflicts with the provider beta filtering behavior.

Confidence Score: 4/5

• Mostly safe to merge after addressing one behavioral bug and aligning documentation with actual provider support.
• Core changes are small and localized, with added tests and a beneficial pricing JSON de-duplication. However, Vertex AI’s new context_management detection will incorrectly emit a beta header for empty/invalid edits, and docs currently claim Bedrock Invoke compaction support that appears inconsistent with the code’s beta filtering.
• litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/experimental_pass_through/transformation.py; docs/my-website/blog/claude_opus_4_6/index.md

Important Files Changed

Filename	Overview
docs/my-website/blog/claude_opus_4_6/index.md	Adds a new section documenting /v1/messages usage and feature matrix; contains a misleading compaction support statement for Bedrock Invoke.
litellm/llms/azure_ai/anthropic/messages_transformation.py	Filters Anthropic beta headers for Azure AI by calling the centralized header filtering helper after updating beta features.
litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py	Replaces inline Bedrock beta whitelist filtering with centralized filter_and_transform_beta_headers(provider='bedrock').
litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/experimental_pass_through/transformation.py	Adds context-management/compaction beta header injection based on optional_params['context_management'].edits; current logic can add context-management header for empty/invalid edits.
litellm/model_prices_and_context_window_backup.json	Removes duplicate claude-opus-4-6-v1/regional pricing entries from backup pricing JSON (de-duplication).
model_prices_and_context_window.json	Removes duplicate claude-opus-4-6-v1/global/regional entries from pricing JSON (fixes duplicate keys that were previously present).
tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_messages_config.py	Adds tests asserting Vertex AI adds compact-2026-01-12 and/or context-management-2025-06-27 to anthropic-beta header when context_management.edits are present.

Sequence Diagram

sequenceDiagram
    participant Client
    participant LiteLLM
    participant AzureAI as Azure AI Anthropic Adapter
    participant Bedrock as Bedrock Anthropic Adapter
    participant Vertex as Vertex AI Anthropic Adapter

    Client->>LiteLLM: POST /v1/messages (Claude 4.6 params)
    alt Azure AI
        LiteLLM->>AzureAI: messages_transformation.py
        AzureAI-->>LiteLLM: provider-formatted payload
    else Bedrock
        LiteLLM->>Bedrock: anthropic_claude3_transformation.py
        Bedrock-->>LiteLLM: provider-formatted payload
    else Vertex AI Partner Model
        LiteLLM->>Vertex: pass_through/transformation.py
        Vertex-->>LiteLLM: provider-formatted payload
    end
    LiteLLM-->>Client: Anthropic-style response

Loading

[greptile-apps]

_{7 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Context management header misfire
When context_management is present but edits is empty (or contains non-dict entries), this logic sets has_other=True for every non-compact_20260112 element (including {} / missing type), which will incorrectly add the context-management-2025-06-27 beta header even though no meaningful edits were requested. Consider only treating an edit as “other” when it’s a dict with a non-empty type (and not equal to compact_20260112).

[greptile-apps]

Incorrect provider support claim
This section states: Compaction is supported on Anthropic, Bedrock Invoke Azure AI, and Vertex AI. but /v1/messages compaction is typically not supported on Bedrock Invoke (it requires anthropic_beta support for the compact feature, which Bedrock Invoke filters/whitelists). Given the code in litellm/llms/bedrock/.../anthropic_claude3_transformation.py explicitly filters provider betas, this doc line is misleading and should be corrected (and also add the missing comma: “Bedrock Invoke, Azure AI”).